Alyssa Ross <hi@alyssa.is> writes:

>> diff --git a/vm/sys/net/xdp-forwarder/default.nix b/vm/sys/net/xdp-forwarder/default.nix
>> new file mode 100644
>> index 0000000..75b1d66
>> --- /dev/null
>> +++ b/vm/sys/net/xdp-forwarder/default.nix
>> @@ -0,0 +1,35 @@
>> +# SPDX-License-Identifier: MIT
>> +# SPDX-FileCopyrightText: 2025 Yureka Lilian <yureka@cyberchaos.dev>
>> +
>> +{ lib, runCommand, stdenv, llvmPackages, libbpf, linux, bpftools }:
>> +
>> +stdenv.mkDerivation {
>> +  pname = "xdp-forwarder";
>> +  version = "0";
>> +
>> +  src = lib.fileset.toSource {
>> +    root = ./.;
>> +    fileset = lib.fileset.fileFilter
>> +      ({ hasExt, ... }: !(hasExt "nix") && !(hasExt "md")) ./.;
>> +  };
>> +
>> +  buildInputs = [ libbpf ];
>> +  nativeBuildInputs = [ llvmPackages.clang-unwrapped bpftools ];
>> +
>> +  buildPhase = ''
>> +    bpftool btf dump file ${linux.dev}/vmlinux format c > include/vmlinux.h
>
> I guess we're still missing a vmlinux.h package in Nixpkgs?  That would
> be much cleaner.

Actually, given discussion in #systemd:nixos.org yesterday, it sounds
like best practice is to declare just the structs and struct members we
use:

https://nakryiko.com/posts/bpf-core-reference-guide/#defining-own-co-re-relocatable-type-definitions

(Since on Spectrum we /can/ predict the kernel we'll be running on, it's
not a huge deal, but I'd still like to follow the principle of having
packages be portable.)